Z.ai and Tsinghua's Tang Jie propose Trace as State, a training-free inference-time scaling method that places reasoning traces before long context, achieving 26 out of 27 wins across models and benchmarks, boosting DeepSeek V4 Pro from 29.2% to 81.8% and GLM-5.2 to 100% on GraphWalks Parents.
The report reveals how Agent competition has shifted from pure model‑parameter races to a full‑system "model+Harness" battle, detailing DeepSeek V4 Pro's technical breakthroughs, massive cost advantage, four‑stage development roadmap, benchmark improvements, industry trends, expert insights, and commercial pathways for AI Agents.
On a single day, Google, DeepSeek, and MiniMax unveiled Gemini 3.7 Flash, V4‑Pro, the Harness runtime, and Music 3, each illustrating how AI is shifting from headline‑grabbing benchmarks toward cost‑effective agents, plug‑in runtimes, and controllable content generation for real‑world workflows.
After integrating the newly released DeepSeek V4 Pro into Codex and running 41 million tokens, the author finds the model’s silent operation, excessive context copying, weak frontend and writing abilities, and higher latency make it unsuitable as a primary model despite solid throughput and low cost.
Kimi K3 is slated for a Q3 release with a massive 2.5 trillion parameters, surpassing DeepSeek V4 Pro and Baidu Wenxin 5.0, reigniting the large‑model arms race and prompting a debate between scale, efficiency, and ecosystem‑driven approaches.
A two‑stage evaluation compares DeepSeek V4 Pro and GPT‑5.3 Codex High on a TypeScript LRU‑Cache task and a markdown‑inspection CLI project, showing DeepSeek leads on basic code correctness while GPT‑5.3 delivers a more complete engineering solution, with detailed scores and analysis.
